Rethinking Tabular Foundation Models On Data Streams

Nilesh Verma; Daniel Nowak-Assis; Afonso Lourenço; Albert Bifet; Bernhard Pfahringer; Maroua Bahri; Nick Jin Sean Lim

cs.LGOct 4, 2026

Rethinking Tabular Foundation Models On Data Streams

Authors: Nilesh Verma, Daniel Nowak-Assis, Afonso Lourenço, Albert Bifet, Bernhard Pfahringer, Maroua Bahri, Nick Jin Sean Lim

Organizations: AI Institute, University of Waikato, New Zealand · LIP6, CNRS, Sorbonne Universit´e, France · GECAD, Polytechnic of Porto, Portugal

Abstract

Tabular foundation models (TFMs) outperform established machine learning models on tabular benchmarks through in-context learning. Building on this success, interest is growing in applying them to data streams, where data arrive continuously and evolve over time. On a stream, a TFM adapts by updating its context rather than its parameters, so its accuracy and cost depend on which examples it keeps and how often it rebuilds its context. We therefore present a systematic study of TFMs on data streams, covering memory management, computational cost, and stream-specific challenges such as concept drift and delayed labels. We find that TFMs achieve the highest predictive performance and that simply retaining the most recent examples is as effective as existing memory management techniques. They also recover faster than streaming learners after drift and keep the highest accuracy under label delay. This accuracy, however, comes at a high serving cost, since a nearly unchanged context is re-encoded at every prediction. These results point to architectural efficiency as the way forward for in-context stream learning.

Figures & tables

Panel of Figure 1 — Figure 1: Approaches to adaptation in data streams, compared under a common prequential protocol measuring predictive performance and computational cost.

Stream	Type	Instances	Features	Classes	Drift
electricity ( Harries, 1999 )	real	45,312	8	2	unknown
noaa ( Elwell and Polikar, 2011 )	real	18,159	8	2	unknown
meter ( Souza et al., 2020b )	real	22,948	96	10	unknown
rialto ( Losing et al., 2016 )	real	82,250	27	10	unknown
posture_no8 ( Kaluza et al., 2010 )	real	163,477	3	10	unknown
agr_a ( Agrawal et al., 1993 )	synth.	30,000	9	2	abrupt

Table 1: Evaluation streams.

Method	mean	med.	mean	med.	std.	mean	rank	mean	rank	mean	rank	mean	med.	mean	med.
	Rank		Accuracy			Balanced acc.		Cohen’s $\kappa$		$\kappa_{t}$		Time (ms/inst.)		Memory (KB/inst.)
TabFM + FIFO	1.11	1	.869	.924	.104	.829	1.33	.772	1.22	.730	1.11	167	129	4.80	1.50
TabICL + FIFO	2.56	2	.860	.906	.105	.820	2.89	.758	2.67	.712	2.56	58.4	71.4	3.73	1.81
CURE	2.89	3	.857	.907	.106	.819	2.89	.754	2.78	.704	2.89	87.3	97.9	5.94	1.96
DR-TabPFN + FIFO	5.33	5	.838	.868	.103	.794	5.67	.724	5.33	.659	5.33	599	542	5.38	1.78
TabPFN + FIFO	6.89	6	.805	.814	.122	.762	7.33	.677	7.44	.619	6.89	103	89.8	2.13	1.32

Table 2: Predictive performance and cost of all methods, averaged over streams. Rank is the per-stream accuracy rank, time uses the geometric mean, and memory is growth per instance. Best in bold , second underlined .

Figure 2: Accuracy over the synthetic streams, each with its own drift type, in stages of roughly a tenth of the stream. Dashes mark concept changes.

Figure 3: Mean accuracy against mean cost over all streams, coloured by method family. Filled marks ( ) are on the Pareto front, hollow marks ( ) are dominated, and grey ties ( ) join each memory management system to FIFO on the same backbone.

Figure 4: Instances to reach the initial plateau (left) and to regain accuracy levels after the first abrupt drift in agr_a (right). A missing bar marks a level not reached before the next drift (dashed line).

Figure 5: Decision functions of kNN and the TFMs on the same FIFO memory shortly after a gradual drift, with the misclassified share of the input area in each title.

Stream	$B$	FIFO	CURE	$\Delta$	ms/inst.	FIFO	DualFIFO	$\Delta$	ms/inst.
		TabICL v2				TabPFN v1
noaa	$100$	.788	.787	$+.001$	21.9	.784	.785	$-.001$	28.3
	$250$	.796	.798	$-.002$	21.0	.792	.785	$+.006$	30.8
	$500$	.814	.813	$+.000$	24.0	.808	.808	$+.000$	41.0
	$1{,}000$	.818	.819	$-.001$	71.8	.814	.814	$+.001$	99.1
agr_a	$100$	.921	.918	$+.003$	17.3	.901	.902	$-.001$	24.9

Table 3: Accuracy and FIFO latency across memory budgets on a real and a synthetic abrupt-drift stream.

\Delta

is FIFO minus the paired memory management policy. Best per stream in bold.

Figure 6: Accuracy of all methods as the label delay

d

grows, where each label arrives

d

instances after its prediction.

Stream	$1$	$10$	$100$	$1{,}000$	$1$	$10$	$100$	$1{,}000$	$1$	$10$	$100$	$1{,}000$
	Context updates				Accuracy				Time (ms/inst.)
noaa	$18{,}059$	$1{,}806$	$181$	$19$	.818	.816	.814	.808	31.5	27.8	27.3	27.1
meter	$22{,}848$	$2{,}285$	$229$	$23$	.906	.903	.875	.710	70.3	62.0	61.3	60.7
agr_a	$29{,}833$	$2{,}984$	$299$	$30$	.935	.934	.932	.917	32.4	28.5	28.0	27.9
tree_r	$59{,}900$	$5{,}990$	$599$	$60$	.767	.768	.764	.737	35.0	31.0	30.7	30.6

Table 4: TabICL with FIFO when the context is refreshed every

K

instances but queried on every instance. Column headers give

K

, and the best per row and metric is in bold.

		Time (ms/inst.)			Accuracy
Backbone	Stream	$q{=}1$	$q{=}10$	$q{=}100$	$q{=}1$	$q{=}10$	$q{=}100$
TabICL v2	meter	88.17	13.08	1.89	.906	.903	.875
	tree_r	110.66	7.69	0.92	.767	.768	.764
TabPFN v1	meter	371.75	50.47	5.58	.736	.735	.709
	tree_r	246.53	11.75	1.52	.646	.646	.644
DR-TabPFN	meter	1367.64	570.62	55.54	.868	.866	.835

Table 5: Query batching with

q

queries sharing one context, for each backbone with FIFO. Best per row and metric in bold.

Panel of Figure 7 — Figure 7: Repeated context encoding versus cached context reuse.

	TabICL v2	TabPFN v1	TabFM
Fixed cost $A$ (ms)	26.1	59.4	120
Per-query cost $\gamma$ ( $\mu$ s)	13.2	22.3	117
Queries per unit of fixed cost $A/\gamma$	$1{,}980$	$2{,}670$	$1{,}030$
Fixed share of a single call	$>99.9\%$	$>99.9\%$	$>99.9\%$

Table 6: Fixed cost

A

and per-query cost

\gamma

fitted to

t(q)=A+q\gamma

on a separate noaa sweep. The lower block is derived from the fit.

Figure 8: Per-instance serving time (left) and peak GPU memory (right) of one compact checkpoint served from a ring cache or by re-encoding every instance, as the memory window grows. The cached path stays flat, whereas re-encoding grows in both time and memory.

Appendix figures & tables24 assets

Supplementary material from the paper’s appendix.

Appendix

Setting	Value	Streams	Where
Default evaluation settings
Warm-up	100 instances	9	all
Evaluation window	1,000, tumbling	9	all
Label delay	0 (immediate)	9	all
Retained memory	1,000 rows	9	all
Queries per prediction	1	9	all

Appendix

Table 7: Default evaluation settings and study-specific variations.

Component	CPU host	GPU hosts
Processor	2 $\times$ AMD EPYC 7702, 128 cores	AMD EPYC 7543, 32 cores
Memory	1 TiB DDR4	251 GiB DDR4
Accelerator	none	NVIDIA RTX A6000, 48 GB
OS	Ubuntu 24.04.4 LTS	Rocky Linux 9.5
Kernel	Linux 6.8.0	Linux 5.14.0
Python	3.14	3.12 (3.11 for the v1 environment)

Appendix

Table 8: Hardware and software used for evaluation.

Method	elec	noaa	meter	rialto	posture	agr_a	sea_g	rbf_m	tree_r	Mean	Rank
No-Change	.853	.680	.011	.000	.203	.647	.561	.260	.505	.413	15.56
Majority-Class	.575	.686	.011	.100	.333	.632	.673	.343	.523	.431	15.22
HAT	.839	.744	.550	.412	.532	.883	.882	.709	.698	.694	12.33
ARF	.900	.798	.683	.728	.619	.875	.890	.888	.738	.791	8.00
SRP	.894	.794	.717	.802	.608	.914	.878	.874	.743	.803	8.11
Lev. Bagging	.890	.785	.618	.612	.600	.821	.886	.861	.691	.752	10.33

Appendix

Table 9: Per-stream accuracy with mean and mean rank across methods, lower rank is better and best values are bold.

	noaa			meter			agr_a			tree_r			Mean			Rank
Method	0	100	1000	0	100	1000	0	100	1000	0	100	1000	0	100	1000	0	100	1000
No-Change	.680	.562	.557	.011	.102	.101	.647	.642	.633	.505	.502	.498	.461	.452	.447	15.8	15.5	15.8
Majority-Class	.686	.686	.686	.011	.100	.101	.632	.632	.632	.523	.523	.523	.463	.485	.486	15.2	15.5	15.2
HAT	.744	.740	.724	.550	.535	.467	.883	.878	.835	.698	.693	.646	.719	.711	.668	11.8	11.8	10.8
ARF	.798	.792	.783	.683	.632	.419	.875	.871	.831	.738	.731	.678	.773	.756	.678	8.8	8.8	9.5
SRP	.794	.791	.782	.717	.668	.446	.914	.907	.858	.743	.737	.682	.792	.776	.692	7.0	7.0	8.0

Appendix

Table 10: Accuracy under label delay

d

on four streams, all sixteen methods. Rank is the mean per-stream rank. Best per column in bold.

Figure 9: Temporal $\kappa_{t}$ across evaluation streams, with methods ordered by median. Fill colour indicates per-instance serving cost on a logarithmic scale.

Figure 10: Mean accuracy ranks with Holm-corrected two-sided Wilcoxon comparisons against the top-ranked method.

Figure 11: Decision functions of kNN and the TFMs on the same FIFO memory after an abrupt drift, a recurrence and a gradual drift, with learned (red dashed) and true (black) boundaries and memory contents as dots.

Figure 12: Decision functions after the first abrupt drift for increasing FIFO memory sizes. Large memories still hold the previous concept.

	$B{=}50$	$100$	$200$	$500$	$1{,}000$	slope
kNN ( $k{=}5$ )	50	50	55	130	271	0.99
TabPFN v1	50	82	109	244	271	0.58
TabICL v2	60	97	203	334	601	0.67
TabFM	60	67	165	335	601	0.80

Appendix

Table 11: Instances after the first drift until windowed accuracy regains

90

percent of its pre-drift level, by FIFO memory size

B

, with the slope on logarithmic axes. The smallest values sit at the measurement floor.

	kNN		TabPFN v1		TabICL v2		TabFM
	all	+300	all	+300	all	+300	all	+300
Best fixed size	.766	.760	.822	.780	.878	.787	.886	.780
$B=50$	.740	.760	.808	.780	.829	.783	.811	.767
$B=1{,}000$	.718	.593	.748	.613	.811	.610	.814	.610
Window oracle	.808	.777	.848	.807	.923	.823	.928	.830
One-switch oracle	.774	.760	.821	.780	.885	.783	.878	.767

Appendix

Table 12: Mean accuracy around the first drift (all) and just after it (+300) for fixed memory sizes, oracles and selectors.

Method	elec	noaa	meter	rialto	posture	agr_a	sea_g	rbf_m	tree_r	Mean	Rank
No-Change	0.20	0.31	0.86	0.19	0.12	0.08	0.10	0.10	0.11	0.17	1.44
Majority-Class	0.30	0.38	0.67	0.16	0.22	0.11	0.09	0.10	0.12	0.19	1.78
HAT	0.31	0.82	1.02	0.35	0.20	0.31	0.17	0.21	0.18	0.32	3.11
ARF	6.50	6.91	41.4	9.46	2.07	6.25	2.49	4.07	6.85	6.30	7.33
SRP	8.87	7.10	53.8	13.3	3.59	8.50	2.89	5.33	10.6	8.47	8.56
Lev. Bagging	2.44	2.96	35.2	6.39	1.64	2.95	1.00	2.64	2.14	3.29	5.78

Appendix

Table 13: Per-stream serving time in ms per instance and mean latency rank, with lower values indicating faster serving.

Method	elec	noaa	meter	rialto	posture	agr_a	sea_g	rbf_m	tree_r	Mean	Rank
No-Change	0.56	2.65	1.22	6.30	1.83	0.07	0.01	0.01	0.07	1.41	4.44
Majority-Class	0.74	2.77	30.3	5.97	0.18	0.18	0.02	0.05	0.05	4.48	5.11
HAT	4.88	3.79	33.6	2.16	0.80	4.27	1.02	1.28	5.27	6.34	9.89
ARF	0.63	6.87	4.83	0.56	4.41	16.6	7.11	10.3	11.1	6.93	11.22
SRP	11.9	4.88	16.1	6.35	5.57	23.4	7.15	6.46	14.8	10.7	14.00
Lev. Bagging	1.09	22.8	2.49	0.24	4.24	6.07	7.08	5.06	2.10	5.69	10.11

Appendix

Table 14: Per-stream resident memory growth in KB per instance and mean rank, with lower values indicating less growth. Dashes mark runs without a usable memory trace; means and ranks use the available streams.

Method	ms/inst.	Host	USD per $10^{6}$ inst.
HAT	0.32	CPU	0.03
SAM-kNN	0.45	CPU	0.05
Lev. Bagging	3.29	CPU	0.35
OAML	3.38	CPU	0.36
ASML	4.55	CPU	0.49
ARF	6.30	CPU	0.67

Appendix

Table 15: Serving cost at on-demand cloud prices, derived from Table 13 . Prices are the AWS m6i.2xlarge rate and the median RTX A6000 rate on 9 September 2026 and are indicative only.

		Windows		Accuracy			$\kappa_{t}$
Stream	prefix	tuning	scored	pretrained	updated	$\Delta$	pretrained	updated	$\Delta$
electricity	$10{,}000$	10	36	.9394	.9388	$-.0006$	$\mathbf{+.5639}$	$+.5600$	$-.0039$
agr_a	$7{,}500$	8	22	.9404	.9406	$+.0002$	$\mathbf{+.6570}$	$+.6566$	$-.0004$

Appendix

Table 16: TabICL performance before and after prefix-based fine-tuning with FIFO memory.

Stream	raw features	TabICL row repr.	TabFM + FIFO
rbf_m	.924	.720	.938
sea_g	.862	.783	.894
agr_a	.838	.805	.942
electricity	.802	.774	.950
noaa	.745	.705	.820
rialto	.618	.539	.924

Appendix

Table 17: Accuracy and serving time for similarity-based prediction using raw features, frozen TabICL representations, and TabFM+FIFO.

	recency decay			drift gate
Stream	off	on	$\Delta$	off	$\Delta$
Real streams
rialto	.540	.618	$-.078$	.659	$+.042$
meter	.503	.527	$-.024$	.534	$+.006$
electricity	.789	.802	$-.013$	.815	$+.013$
noaa	.743	.745	$-.002$	.744	$-.001$

Appendix

Table 18: Ablation of recency decay and the drift gate in the raw-feature similarity predictor.

\Delta

is off minus on, so negative values favour the component.

Panel of Figure 13 — Figure 13: Datapoint-axis attention masks for bidirectional, causal, and sliding attention.

Panel of Figure 14 — Figure 14: Cache dependencies in the evaluated TabICL-style architecture.

Figure 15: Accuracy sensitivity to the sliding-attention window size.

	nanoTabPFN, $356$ k			nanoTabICL, $8.1$ M
Drift type	bidirectional	causal	sliding	bidirectional	causal	sliding
abrupt	.715	.880	.891	.749	.793	.784
gradual	.687	.876	.889	.743	.758	.755
incremental	.648	.819	.856	.761	.781	.780
recurring	.679	.620	.623	.737	.747	.747
stationary	.883	.875	.900	.875	.874	.870

Appendix

Table 19: Accuracy across datapoint-axis attention masks and drift conditions.

Figure 16: Accuracy across bidirectional, causal, and sliding attention masks.

$W$	abrupt	gradual	incremental	recurring	stationary	mean
$32$	.883	.896	.863	.748	.869	.852
$64$	.901	.913	.877	.750	.887	.866
$128$	.904	.916	.876	.732	.896	.865
$256$	.906	.912	.865	.770	.910	.872
$512$	.897	.917	.863	.776	.914	.874
bidirectional	.804	.831	.776	.789	.914	.823

Appendix

Table 20: Accuracy across sliding-attention window sizes for nanoTabPFN, trained separately from Table 19 .

Model	acc.	$\kappa_{t}$	acc.	$\kappa_{t}$	acc.	$\kappa_{t}$	acc.	$\kappa_{t}$
	electricity		noaa		sea_g		tree_r
No-Change	.853	$-$ .000	.680	$-$ .000	.561	$-$ .000	.505	$-$ .000
TabICL + FIFO	.944	$+.616$	.818	$+.431$	.891	$+.753$	.767	$+.530$
sliding	.602	$-1.71$	.732	$+.160$	.814	$\mathbf{+.575}$	.559	$\mathbf{+.108}$
sliding, learned recency bias	.428	$-2.90$	.742	$\mathbf{+.194}$	.732	$+.389$	.551	$+.091$

Appendix

Table 21: Transfer accuracy and temporal

\kappa_{t}

on evaluation streams. Bold marks the better compact model.

Stream	$B$	Cached	Re-enc.	$\Delta$	Cached	Re-enc.	Cached	Re-enc.	Cached	Re-enc.
		Accuracy			$\kappa_{t}$		Time (ms/inst.)		Peak GPU (MB)
electricity	$128$	.609	.395	$+.214$	$-1.469$	$-2.824$	8.2	5.6	26.5	31.3
	$250$	.609	.609	$.000$	$-1.469$	$-1.469$	7.7	5.7	26.5	62.7
	$500$	.609	.404	$+.205$	$-1.469$	$-2.768$	8.1	10.4	26.5	173
	$1{,}000$	.609	.405	$+.205$	$-1.469$	$-2.764$	8.2	28.0	26.5	604
	$2{,}000$	.609	.430	$+.179$	$-1.469$	$-2.605$	8.2	83.1	26.5	2,297

Appendix

Table 22: The same sliding-window checkpoint served from cache and by re-encoding on a stream prefix, per memory budget.

\Delta

is cached minus re-encoded accuracy, time is per instance, and peak GPU memory is the allocator high-water mark, which depends only on the feature count and

B

. Electricity was timed on a second host of the same type.

Explore similar work

Bounded Context Management for Tabular Foundation Models on Stream Learning
Jun 17, 2026Jinmo Lee, Doyun Choi, Moongi Choi +1Tabular Foundation Models Tabular Learning
Department of Computer Science and Engineering, Seoul National University, Seoul, Republic of Korea · The Kim Jaechul Graduate School of AI, KAIST, Daejeon, Republic of Korea.
Memory Efficient Tabular Foundation Models
Jul 30, 2026Shuting Luo, Monika Mikhail Kanaan, Cameron Gordon +2Tabular Foundation Models Tabular Learning
Commonwealth Bank of Australia, Australia · Australian Institute for Machine Learning, University of Adelaide, Adelaide, Australia
Streaming Hierarchical Inference with Tabular Foundation Models
Sep 7, 2026Vitor Crista, Afonso Lourenço, Diogo Martinho +1Tabular Foundation Models Probabilistic Inference
GECAD, Polytechnic of Porto, Portugal

Jun 17, 2026, cs.LGJ/K move · Enter open · S save

Bounded Context Management for Tabular Foundation Models on Stream Learning

Jinmo Lee, Doyun Choi, Moongi Choi, Jaemin Yoo

Department of Computer Science and Engineering, Seoul National University, Seoul, Republic of Korea · The Kim Jaechul Graduate School of AI, KAIST, Daejeon, Republic of Korea.

Tabular Foundation Models Tabular Learning In-Context Learning Replay-Based Continual Learning

Open paper AskPDF arXiv

Tabular stream learning requires predictions on sequentially arriving examples under distribution shift. While standard methods adapt by updating model states, tabular foundation models (TFMs) make predictions conditioned on a labeled context in an in-context manner, making them a natural alternative for stream learning. This shifts the challenge from how to update the model to how to manage the context. We propose a future information view that yields three practical requirements for context management: preserve recent examples, retain uncertain examples, and remove redundant examples. We instantiate these requirements as CURE (Context management via Uncertainty-aware admission and Redundancy aware Eviction), a context-managing policy with entropy-gated admission and redundancy-aware eviction. Across seven streams, CURE shows up to 27.0% relative improvement over classical stream learners, remains robust across multiple TFM backbones, and ranks first among other policy variants. Code and datasets are available at https://github.com/morcellinus/CURE-ICML-FMSD.

Abstract

Explore similar work

Bounded Context Management for Tabular Foundation Models on Stream Learning

Memory Efficient Tabular Foundation Models

Streaming Hierarchical Inference with Tabular Foundation Models